Robotics 2

Position Regulation
(with an introduction to stability)

Prof. Alessandro De Luca

 Equilibrium states of a robot
𝑥# 𝑞
𝑀 𝑞 𝑞̈ + 𝑐 𝑞, 𝑞̇ + 𝑔 𝑞 = 𝑢 𝑥 = 𝑥 = 𝑞̇
$

𝑥̇ # 𝑥$ 0
𝑥̇ = = −𝑀/# (𝑥 ) 𝑐 𝑥 , 𝑥 + 𝑔(𝑥 ) + 𝑢
𝑥̇ $ # # $ # 𝑀/# (𝑥# )
= 𝑓 𝑥 + 𝐺 𝑥# 𝑢

𝑥5 unforced equilibrium 𝑥5,$ = 0

𝑓 𝑥5 = 0
(𝑢 = 0) 𝑔 𝑥5,# = 0
𝑥5 forced equilibrium 𝑥5,$ = 0
𝑓 𝑥5 + 𝐺 𝑥5,$ 𝑢(𝑥5 ) = 0
(𝑢 = 𝑢(𝑥)) 𝑢 𝑥5 = 𝑔 𝑥5,#
all equilibrium states of mechanical joint torques must balance gravity
systems have zero velocity! at the equilibrium!
Robotics 2 2
 Stability of dynamical systems
definitions - 1

𝑥̇ = 𝑓(𝑥) e.g., a closed-loop system

(under feedback control)
𝑥5 equilibrium: 𝑓(𝑥5 ) = 0
(sometimes we consider as equilibrium state
𝑥5 = 0, e.g., when using errors as variables)

stability of 𝑥5
∀𝜀 > 0, ∃𝛿; > 0: 𝑥 𝑡> − 𝑥5 < 𝛿; ⟹ 𝑥 𝑡 − 𝑥5 < 𝜀, ∀𝑡 ≥ 𝑡>

asymptotic stability of 𝑥5 stability +

∃𝛿 > 0: 𝑥 𝑡> − 𝑥5 < 𝛿 ⟹ 𝑥 𝑡 − 𝑥5 → 0, for 𝑡 → ∞

asymptotic stability may become global (∀𝛿 > 0, finite)

note: these are definitions of stability “in the sense of Lyapunov”

Robotics 2 3
 Stability vs. asymptotic stability
whiteboard…

𝑥5 𝛿; 𝑥5 𝑥5 𝑥5
𝜀
𝑥(𝑡> ) 𝑥(𝑡> )

∀𝜀 > 0 ∃𝛿; > 0 𝑥 𝑡> − 𝑥5 < 𝛿; ⟹ 𝑥 𝑡 − 𝑥5 < 𝜀, ∀𝑡 ≥ 𝑡>

equilibrium state 𝑥5 is stable

𝑥5 𝑥5 𝑥5
+
𝛿 𝑥(𝑡> ) 𝑥(𝑡> )

∃𝛿 > 0 𝑥 𝑡> − 𝑥5 < 𝛿 ⟹ 𝑥 𝑡 − 𝑥5 → 0, for 𝑡 → ∞

equilibrium state 𝑥5 is asymptotically stable

Robotics 2 4
 Stability of dynamical systems
definitions - 2
exponential rate 𝜆
exponential stability of 𝑥5
∃𝛿, 𝑐, 𝜆 > 0: 𝑥 𝑡> − 𝑥5 < 𝛿 ⟹ 𝑥 𝑡 − 𝑥5 ≤ 𝑐𝑒 /M N/NO 𝑥 𝑡> − 𝑥5

§ allows to estimate the time needed to ”approximately” converge: for 𝑐 = 1,

in 𝑡 − 𝑡> = 3 × the time constant 𝜏 = 1/𝜆, the initial error is reduced to 5%
§ typically, this is a local property only (within some maximum finite radius 𝛿 )
⇒ such “domain of attraction” is hard to be estimated accurately

“practical” stability of a set 𝑆

∃𝑇 𝑥 𝑡> , 𝑆 ∈ ℝ: 𝑥 𝑡 ∈ 𝑆, ∀𝑡 ≥ 𝑡> + 𝑇 𝑥 𝑡> , 𝑆

a finite time also known as u.u.b. stability

⇒ trajectories 𝑥(𝑡) are “ultimately uniformly bounded” (use in robust control)
Robotics 2 5
 The need for analysis and criteria
whiteboard…

a nonlinear system 𝑥̇ = 𝑓 𝑥 in ℝ$ two equilibria 𝑓 𝑥5 = 0

𝑥̇ # = 1 − 𝑥#U
T 𝑥5V = 1, 1 , 𝑥5VV = 1, −1
𝑥̇ $ = 𝑥# − 𝑥$$
to assess (asymptotic) stability [or not]
1.5
asymptotically stable 𝑥5V

of equilibria, do we need to compute 1

all system trajectories, starting from all 𝑥(𝑡> )

possible initial states 𝑥(𝑡> )?
0.5

x2
0

rather, we may be able to just look at

-0.5

the time evolution of a scalar function 𝑉, -1

𝑥(𝑡> )
evaluated analytically along the state unstable 𝑥5VV
trajectories of the system (even in ℝX !)
-1.5
-1 -0.5 0 0.5 1 1.5
x1

Robotics 2 6
 Stability of dynamical systems
results - 1

Lyapunov candidate 𝑉 𝑥 : ℝX ⟶ ℝ such that

positive
𝑉 𝑥5 = 0, 𝑉 𝑥 > 0, ∀𝑥 ≠ 𝑥5 definite
function
#
typically quadratic (e.g., (𝑥 − 𝑥5 )Y 𝑃(𝑥 − 𝑥5 ) with level surfaces = ellipsoids)
$
may also be a local candidate only (∀𝑥 ≠ 𝑥5 : 𝑥 − 𝑥5 < 𝛿)
sufficient condition of stability
negative
̇
∃𝑉 candidate: 𝑉(𝑥) ≤ 0, along the trajectories of 𝑥̇ = 𝑓(𝑥) semi-definite
function

sufficient condition of asymptotic stability negative

∃𝑉 candidate: 𝑉̇ 𝑥 < 0, along the trajectories of 𝑥̇ = 𝑓(𝑥) definite
function

sufficient condition of instability

∃𝑉 candidate: 𝑉̇ 𝑥 > 0, along the trajectories of 𝑥̇ = 𝑓(𝑥)
Robotics 2 7
 Stability of dynamical systems
results - 2

sufficient condition of u.u.b. stability of a set 𝑆

∃𝑉 candidate: i) 𝑆 is a level set of 𝑉 for a given 𝑐>
𝑆 = 𝑆 𝑐> = 𝑥 ∈ ℝX : 𝑉(𝑥) ≤ 𝑐>
ii) 𝑉̇ 𝑥 < 0 along trajectories of 𝑥̇ = 𝑓 𝑥 , 𝑥 ∉ 𝑆

LaSalle Theorem
if ∃𝑉 candidate: 𝑉̇ 𝑥 ≤ 0 along the trajectories of 𝑥̇ = 𝑓(𝑥)

then system trajectories asymptotically converge to the

̇
largest invariant set ℳ ⊆ 𝑆 = 𝑥 ∈ ℝX : 𝑉(𝑥) =0
ℳ is invariant if 𝑥 𝑡> ∈ ℳ ⟹ 𝑥 𝑡 ∈ ℳ, ∀𝑡 ≥ 𝑡>

Corollary
ℳ ≡ 𝑥5 asymptotic stability
Robotics 2 8
 Bird-eye view on Lyapunov analysis
whiteboard…

a mass 𝑚 at the end of an unforced (passive) pendulum of length 𝑙

𝑚𝑙 $ 𝜃̈ + 𝑏 𝜃̇ + 𝑚𝑙𝑔> sin 𝜃 = 0 𝑥 = 𝑥# , 𝑥$ 𝑥̇ # = 𝑥$
⇒ ̇ ∈ ℝ$
⇒ 𝑔> 𝑏
lower equilibrium at 𝜃5 = 0 = (𝜃, 𝜃) 𝑥̇ $ = − sin 𝑥# − 𝑥
𝑙 𝑚𝑙 $ $
4

2
2

1
1

phase 𝑥5 = 𝟎 𝑥5 = 𝟎
dtheta

dtheta
0 0

plane -1
-1

-2
-2

-3 level
-4
sets of 𝑉 -3

-4 -3 -2 -1 0 1 2 3 4 -3 -2 -1 0 1 2 3

1
theta theta

𝑉 = 𝐸 = 𝑚𝑙 $ 𝜃̇ $ + 𝑚𝑙𝑔> (1 − cos 𝜃) ≥ 0 𝑉 = 0 ⇔ 𝑥5 = 𝜃5 , 𝜃̇5 = (0,0)

2
𝑉̇ = 𝜃̇ 𝑚𝑙 $ 𝜃̈ + 𝑚𝑙𝑔> sin 𝜃 = −𝑏 𝜃̇ $ ≤ 0 stability of equilibrium 𝑥5 = 0
⇒ (… at least!)
nO
⇒ use LaSalle: 𝑉̇ = 0 ⇔ 𝜃̇ = 0 ⇒ 𝜃̈ = − o
sin 𝜃 ≠ 0 unless 𝜃 = 𝜃5 = 0 (or 𝜋 !)
Robotics 2 ⇒ local asymptotic stability 9
 Stability of dynamical systems
results - 3

§ previous results are also valid for periodic time-varying systems

𝑥̇ = 𝑓 𝑥, 𝑡 = 𝑓 𝑥, 𝑡 + 𝑇t ⇒ 𝑉 𝑥, 𝑡 = 𝑉 𝑥, 𝑡 + 𝑇t
§ for general time-varying systems (e.g., in robot trajectory tracking control)

𝑥̇ = 𝑓(𝑥, 𝑡)
Barbalat Lemma
if i) a function 𝑉 𝑥, 𝑡 is lower bounded
ii) 𝑉̇ 𝑥, 𝑡 ≤ 0
then ⇒ ∃ lim 𝑉 𝑥, 𝑡 (but this does not imply that lim 𝑉̇ 𝑥, 𝑡 = 0)
N→s N→s
if in addition iii) 𝑉̈ 𝑥, 𝑡 is bounded
then ⇒ lim 𝑉̇ 𝑥, 𝑡 = 0
N→s

Corollary
if a Lyapunov candidate 𝑉(𝑥, 𝑡) satisfies Barbalat Lemma along the
trajectories of 𝑥̇ = 𝑓(𝑥, 𝑡), then the conclusions of LaSalle Theorem hold
Robotics 2 10
 Stability of linear systems
time-invariant case

𝑥̇ = 𝐴𝑥 𝑥5 = 0 is always an equilibrium state

I. asymptotic stability
II. global asymptotic stability
III. exponential stability
IV. 𝜎(𝐴) ⊂ ℂ/ (all eigenvalues of 𝐴 have negative real part)
V. ∀𝑄 > 0 (positive definite), ∃! 𝑃 > 0: 𝐴Y 𝑃 + 𝑃𝐴 = −𝑄
# Y
Lyapunov equation 𝑥 𝑃𝑥 is a Lyapunov candidate
$
ALL EQUIVALENT !!
if 𝑥5 = 0 is an asymptotically stable equilibrium,
then it is necessarily the unique equilibrium
Robotics 2 11
 Stability of the linear approximation

|}
̇
Let ∆𝑥 = 𝑥 − 𝑥5 and let ∆𝑥 = | 𝑥 − 𝑥5 = 𝐴 ∆𝑥 be the
|~ ~€~•
linear approximation of 𝑥̇ = 𝑓(𝑥) around the equilibrium 𝑥5

𝐴 asymptotically stable (𝜎(𝐴) ⊂ ℂ/)

the original nonlinear system is

exponentially stable at the origin

this is only a local result

(used also to estimate the domain of attraction)

Robotics 2 12
 PD control
(proportional + derivative action on the error)

robot 𝑀 𝑞 𝑞̈ + 𝑐 𝑞, 𝑞̇ + 𝑔 𝑞 = 𝑢

goal: asymptotic stabilization (= regulation)

of the closed-loop equilibrium state
𝑞 = 𝑞| , 𝑞̇ = 0

possibly obtained from kinematic inversion: 𝑞| = 𝑓 /#(𝑟| )

control law 𝑢 = 𝐾ƒ 𝑞| − 𝑞 − 𝐾„ 𝑞̇

𝐾ƒ > 0, 𝐾„ > 0 (positive definite), symmetric

Robotics 2 13
 Asymptotic stability with PD control
Theorem 1
In the absence of gravity (𝑔(𝑞) = 0), the robot state (𝑞| , 0) under
the given PD joint control law is globally asymptotically stable

Proof let 𝑒 = 𝑞| − 𝑞 (𝑞| constant)

1 Y 1 Y
Lyapunov candidate 𝑉 = 𝑞̇ 𝑀 𝑞 𝑞̇ + 𝑒 𝐾ƒ 𝑒 ≥ 0 𝑉 = 0 ⇔ 𝑒 = 𝑒̇ = 0
2 2
# #
𝑉̇ = 𝑞̇ Y 𝑀 𝑞̈ + 𝑞̇ Y 𝑀̇ 𝑞̇ − 𝑒 Y 𝐾ƒ 𝑞̇ = 𝑞̇ Y 𝑢 − 𝑆 𝑞̇ + 𝑀̇ 𝑞̇ − 𝑒 Y 𝐾ƒ 𝑞̇
$ $

= 0, due to energy conservation

= 𝑞̇ Y 𝐾ƒ 𝑒 − 𝑞̇ Y 𝐾„ 𝑞̇ − 𝑒 Y 𝐾ƒ 𝑞̇ = −𝑞̇ Y 𝐾„ 𝑞̇ ≤ 0 (𝐾„ > 0, symmetric)

up to here, we proved
stability only
but 𝑉̇ = 0 ⇔ 𝑞̇ = 0 continues ...

Robotics 2 14
 Asymptotic stability with PD control
LaSalle
system trajectories converge to the largest
𝑉̇ = 0 ⇔ 𝑞̇ = 0
invariant set of states ℳ where 𝑞̇ ≡ 0
(that is 𝑞̇ = 𝑞̈ = 0)

𝑞̇ = 0 𝑀 𝑞 𝑞̈ = 𝐾ƒ 𝑒 𝑞̈ = 𝑀/# (𝑞)𝐾ƒ 𝑒
closed-loop dynamics invertible

𝑞̇ = 0, 𝑞̈ = 0 ⇔ 𝑒 = 0

the only invariant state in 𝑉̇ = 0 is given by 𝑞 = 𝑞| , 𝑞̇ = 0

note: typically, 𝐾ƒ = diag 𝑘ƒŠ , 𝐾„ = diag 𝑘„Š ,

decentralized linear control (local to each joint)

Robotics 2 15
 Mechanical interpretation
n for diagonal positive definite gain matrices 𝐾ƒ and 𝐾„ (thus, with
positive diagonal elements), such values correspond to stiffness of
“virtual” springs and viscosity of “virtual” dampers placed at the joints

stiffness 𝑘ƒŠ > 0 𝑞| (single components

are defined relative
viscosity 𝑘„Š > 0 to the previous link!!)

current
configuration 𝑞
desired
configuration 𝑞|

Robotics 2 16
 Plot of the Lyapunov function 𝑉
n time evolution of the Lyapunov candidate
𝑉(0) = ½ 𝑒 Y (0)𝐾ƒ 𝑒(0)
𝑉 if the robot starts from rest
𝑉(0)
𝑉̇ = 0

0
(isolated) instants of global “motion inversion”
𝑡
(𝑞̇ = 0, but 𝑞̈ ≠ 0!)

Robotics 2 17
 Comments on PD control -1

n choice of control gains affects robot evolution during transients and

practical settling times
n hard to define values that are “optimal” in the whole workspace

n “full” 𝐾ƒ and 𝐾„ gain matrices allow to assign desired eigenvalues

to the linear approximation of the robot dynamics around the final
desired state (𝑞| , 0)
n when (joint) viscous friction is present, the derivative term in the
control law is not strictly necessary
n −𝐹• 𝑞̇ in the robot model acts similarly to −𝐾„ 𝑞̇ in the control law,
but the latter can be modulated at will
n in the absence of tachometers, the actual realization of the derivative
term in the feedback law requires some processing of joint position
data measured by digital encoders (or analog resolvers/potentiometers)

Robotics 2 18
 Comments on PD control -2
n analog or digital implementation of derivative action in the control law
when only position is measured at the joints (e.g., through encoders)
continuous-time
𝑢 𝑡 = 𝐾ƒ 𝑒 𝑡 + 𝐾„ 𝑒(𝑡)
̇ 𝑒 = 𝑞| − 𝑞, 𝑒̇ = −𝑞̇
control law (design)

representation in 𝐾„ 𝑠
𝑢 𝑠 = 𝐾ƒ + 𝐾„ 𝑠 𝑒(𝑠) 𝑢 𝑠 = 𝐾ƒ + 𝑒(𝑠)
the Laplace domain 1 + 𝜏𝑠
not realizable as such
(non-proper transfer function) derivative action limited
in bandwidth (up to ω ≤ 1/𝜏)

transformation in 1 − 𝑧 /#
the Zeta-domain 1 − 𝑧 /# 𝑇•
(e.g., via backward 𝑢 𝑧 = 𝐾ƒ + 𝐾„ 𝑒(𝑧) 𝑢 𝑧 = 𝐾ƒ + 𝐾„ 𝑒(𝑧)
𝑇• 1 − 𝑧 /#
differentiation rule on 1+𝜏
samples, every 𝑇• sec) 𝑇•

discrete-time 𝑒’ − 𝑒’/# 𝐾„
𝑢’ = 𝐾ƒ 𝑒’ + 𝐾„ 𝑢’ = 𝐾ƒ 𝑒’ + 𝑒’ − 𝑒’/#
implementations 𝑇• 𝜏 + 𝑇•
𝜏
both realizable + (𝑢’/# − 𝐾ƒ 𝑒’/# )
Robotics 2
𝜏 + 𝑇• 19
 Inclusion of gravity
n in the presence of gravity, the same previous arguments
(and proof) show that the control law
𝑢 = 𝐾ƒ 𝑞| − 𝑞 − 𝐾„ 𝑞̇ + 𝑔(𝑞) 𝐾ƒ > 0, 𝐾„ > 0
will make the equilibrium state (𝑞| , 0) globally
asymptotically stable (nonlinear cancellation of gravity)
§ if gravity is not cancelled or only approximately cancelled
𝑢 = 𝐾ƒ 𝑞| − 𝑞 − 𝐾„ 𝑞̇ + 𝑔(𝑞)
” 𝑔(𝑞)
” ≠ 𝑔(𝑞)
it is 𝑞 → 𝑞∗ ≠ 𝑞| , 𝑞̇ → 0, with steady-state position error
n 𝑞 ∗ is not unique in general, except when 𝐾ƒ is chosen large enough
n explanation in terms of linear systems: there is no integral action before
the point of access of the constant “disturbance” acting on the system

Robotics 2 20
 PD control +
constant gravity compensation
since 𝑔(𝑞) contains only trigonometric and/or linear terms in 𝑞,
the following structural property holds

𝜕$𝑈 𝜕𝑔
finite ∃𝛼 > 0: $
= ≤ 𝛼, ∀𝑞
𝜕𝑞 𝜕𝑞

consequence 𝑔 𝑞 − 𝑔(𝑞| ) ≤ 𝛼 𝑞 − 𝑞|
induced
note: norm of 𝐴 = 𝜆•–— 𝐴Y 𝐴 ≜ 𝐴™ ≥ 𝐴š ≜ 𝜆•›œ(𝐴Y 𝐴)
a matrix

LINEAR CONTROL law

𝐾ƒ , 𝐾„ > 0
𝑢 = 𝐾ƒ 𝑞| − 𝑞 − 𝐾„ 𝑞̇ + 𝑔(𝑞| ) symmetric

linear feedback + constant feedforward

Robotics 2 21
 PD control + constant gravity compensation
stability analysis

Theorem 2
If 𝐾ƒ,š > 𝛼 , the state (𝑞| , 0) of the robot under joint-space PD control
+ constant gravity compensation at 𝑞| is globally asymptotically stable

Proof
1. (𝑞| , 0) is the unique closed-loop equilibrium state

in fact, for 𝑞̇ = 0 and 𝑞̈ = 0, it is 𝐾ƒ 𝑒 = 𝑔(𝑞) − 𝑔(𝑞| )

which can hold only for 𝑞 = 𝑞| , because when 𝑞 ≠ 𝑞|

𝐾ƒ 𝑒 ≥ 𝐾ƒ,š 𝑒 > 𝛼 𝑒 ≥ 𝑔 𝑞 − 𝑔(𝑞| )

Robotics 2 22
 PD control + constant gravity compensation
stability analysis

¡ Y
with 𝑒 = 𝑞| − 𝑞, 𝑔 𝑞 = , consider as Lyapunov candidate
¢
1 Y 1
𝑉= 𝑞̇ 𝑀 𝑞 𝑞̇ + 𝑒 Y 𝐾ƒ 𝑒 + 𝑈 𝑞 − 𝑈 𝑞| + 𝑒 Y 𝑔(𝑞| )
2 2

2. 𝑉 is convex in 𝑞̇ and 𝑒, and zero only for 𝑒 = 𝑞̇ = 0

• Y
= 𝑀 𝑞 𝑞̇ = 0 only for 𝑞̇ = 0 (𝑞| , 0) is a
¢̇
£• global minimum
=𝑀 𝑞 >0 of 𝑉 ≥ 0
¢̇ £
Y
•|¤¥O
̇ ¡ Y
= 𝐾ƒ 𝑒 − + 𝑔 𝑞| = 𝐾ƒ 𝑒 + 𝑔 𝑞| − 𝑔 𝑞 = 0
5 ¢
𝜕𝑒 ⁄𝜕𝑞 = −𝐼 only for 𝑞 = 𝑞|
£• £¡
̇
|¤¥O
= 𝐾ƒ + > 0, since 𝐾ƒ = 𝐾ƒ,™ ≥ 𝐾ƒ,š > 𝛼
5£ ¢£

Robotics 2 23
 PD control + constant gravity compensation
stability analysis
1 Y 1 Y
differentiating 𝑉 = 𝑞̇ 𝑀 𝑞 𝑞̇ + 𝑒 𝐾ƒ 𝑒 + 𝑈 𝑞 − 𝑈 𝑞| + 𝑒 Y 𝑔(𝑞| )
2 2
1 𝜕𝑈(𝑞)
̇ Y ̇ Y
𝑉 = 𝑞̇ 𝑀(𝑞)𝑞̈ + 𝑀(𝑞)𝑞̇ − 𝑒 𝐾ƒ 𝑞̇ + 𝑞̇ − 𝑞̇ Y 𝑔(𝑞| )
2 𝜕𝑞
1
Y
= 𝑞̇ 𝑢 − 𝑆(𝑞, 𝑞) ̇
̇ 𝑞̇ + 𝑀(𝑞) 𝑞̇ − 𝑔(𝑞) − 𝑒 Y 𝐾ƒ 𝑞̇ + 𝑞̇ Y 𝑔 𝑞 − 𝑔(𝑞| )
2
=0
= 𝑞̇ Y 𝐾ƒ 𝑒 − 𝑞̇ Y 𝐾„ 𝑞̇ + 𝑞̇ Y 𝑔 𝑞| − 𝑔(𝑞) − 𝑒 Y 𝐾ƒ 𝑞̇ + 𝑞̇ Y 𝑔 𝑞 − 𝑔(𝑞| )

= −𝑞̇ Y 𝐾„ 𝑞̇ ≤ 0

for 𝑉̇ = 0 (⇔ 𝑞̇ = 0), we have in the closed-loop system

𝑀 𝑞 𝑞̈ + 𝑔 𝑞 = 𝐾ƒ 𝑒 + 𝑔(𝑞| )
𝑞̈ = 𝑀 /# 𝑞 𝐾ƒ 𝑒 + 𝑔 𝑞| − 𝑔 𝑞 =0⇔𝑒=0

by LaSalle Theorem, the thesis follows

Robotics 2 24
 Example of a single-link robot
stability analysis
task: regulate the link position to the upward equilibrium
𝜃| = 𝜋 → 𝑔 𝜃| = 0
PD control + constant gravity compensation (here, zero!)
𝑢 = 𝑘ƒ 𝜋 − 𝜃 − 𝑘„ 𝜃̇
by Theorem 2, it is sufficient (here, also necessary*) to choose
𝑘ƒ > 𝛼 = 𝑚𝑔> 𝑑, 𝑘„ > 0
𝐼𝜃̈ + 𝑚𝑔> 𝑑 sin 𝜃 = 𝑢
single
minimum
two local minima
in 𝜃| = 𝜋

rad
𝑘ƒ 𝑘ƒ 𝑘ƒ
©𝑚𝑔 𝑑 = 0.5 ©𝑚𝑔 𝑑 = 1 ©𝑚𝑔 𝑑 = 5
> > >

plots of 𝑉(𝜃) (for 𝜃̇ = 0) * by a local analysis of the

Robotics 2 linear approximation at 𝜋 25
 Example of a single-link robot
simulations with data: 𝐼 = 0.9333, 𝑚𝑔>𝑑 = 19.62 (= 𝛼)
𝜃| = 180° → 𝑔 𝜃| = 0 𝜃| = 90° → 𝑔 𝜃| = 𝑚𝑔> 𝑑
sufficient P gain: 𝑘ƒ = 36, 𝑘„ = 12 low P gain: 𝑘ƒ = 16, 𝑘„ = 8 low P gain: 𝑘ƒ = 16, 𝑘„ = 8
position
velocity
error

o
zero final error 60 residual error no residual error!

large initial torque small initial torque even smaller initial torque
control

Robotics 2 26
 Approximate gravity compensation

the approximate control law

𝑢 = 𝐾ƒ 𝑞| − 𝑞 − 𝐾„ 𝑞̇ + 𝑔(𝑞
” |)

leads, under similar hypotheses, to a closed-loop equilibrium 𝑞∗

§ its uniqueness is not guaranteed (unless 𝐾ƒ is large enough)

§ for 𝐾ƒ → ∞, one has 𝑞∗ → 𝑞|

Conclusion: In the presence of gravity, the previous regulation control laws

require an accurate knowledge of the gravity term in the dynamic model
in order to guarantee the zeroing of the position error
(since we can only use “finite” control gains ⇒ in practice, not too large)

Robotics 2 27
 PID control
n in linear systems, the addition of an integral control action is used
to eliminate a constant error in the step response at steady state
n in robots, a PID may be used to recover such a position error due
to an incomplete (or absent) gravity compensation/cancellation
N
the control law 𝑢 𝑡 = 𝐾ƒ 𝑞| − 𝑞(𝑡) + 𝐾± ² 𝑞| − 𝑞(𝜏) 𝑑𝜏 − 𝐾„ 𝑞(𝑡)
̇
>

n is independent from any robot dynamic model term

n if the desired closed-loop equilibrium is asymptotically stable under
PID control, the integral term is “loaded” at steady state to the value
s
𝐾± ² 𝑞| − 𝑞(𝜏) 𝑑𝜏 = 𝑔(𝑞| )
>
n however, one can show only local asymptotic stability of this law, i.e.,
for 𝑞(0) ∈ Δ(𝑞| ), under complex conditions on 𝐾ƒ , 𝐾± , 𝐾„ and 𝑒(0)
Robotics 2 28
 Linear example with PID control
whiteboard…

𝑒 𝑡 = 𝑞| − 𝑞 𝑡
𝑚𝑞̈ + 𝑚𝑔> = 𝐹 (no friction)
𝑔> > 0 𝑒̇ 𝑡 = −𝑞̇ 𝑡
𝑚 𝐹 = 𝑘ƒ 𝑞| − 𝑞 − 𝑘„ 𝑞̇
𝑞
šnO
(PD ⇒ steady-state error 𝑒 = 𝑞| − 𝑞
³ , with 𝑞³ = 𝑞| − ’ )
´

𝐹 𝐹 = 𝑘ƒ 𝑞| − 𝑞 − 𝑘„ 𝑞̇ + 𝑚𝑔>
(PD + gravity cancellation ⇒ regulation ∀𝑘ƒ > 0, 𝑘„ > 0)
N
𝐹 = 𝑘ƒ 𝑞| − 𝑞 − 𝑘„ 𝑞̇ + 𝑘± ² 𝑞| − 𝑞(𝜏) 𝑑𝜏
> š’µ
(PID ⇒ regulation ∀𝑘± > 0, 𝑘„ > 0, 𝑘ƒ > > 0)
’¶
Laplace domain analysis: 𝑒 𝑠 = ℒ[𝑒 𝑡 ], 𝑑 𝑠 = ℒ[𝑚𝑔>] + Routh criterion

𝑒(𝑠) 𝑠 3 𝑚 𝑘ƒ
= 𝑊| 𝑠 = 2 𝑘„ 𝑘±
𝑑(𝑠) 𝑚𝑠 U + 𝑘„ 𝑠 $ + 𝑘ƒ 𝑠 + 𝑘±
1 (𝑘„ 𝑘ƒ − 𝑚𝑘± )/𝑘„
Robotics 2 0 𝑘± 29
 Saturated PID control
§ more in general, one can prove global asymptotic stability of (𝑞| , 0),
under lower bound limitations for 𝐾ƒ , 𝐾± , 𝐾„ (depending on suitable
“bounds” on the terms in the dynamic model), for a nonlinear PID law
N
𝑢 𝑡 = 𝐾ƒ 𝑞| − 𝑞(𝑡) + 𝐾± ² Φ 𝑞| − 𝑞(𝜏) 𝑑𝜏 − 𝐾„ 𝑞̇
>

where Φ(𝑞| − 𝑞) is a saturation-type function, such as

sin 𝑥 , 𝑥 ≤ 𝜋/2 𝑒 ~ − 𝑒 /~
Φ 𝑥 = ¾ 1, 𝑥 > 𝜋/2 or Φ 𝑥 = tanh 𝑥 = ~
−1, 𝑒 + 𝑒 /~
𝑥 < −𝜋/2

(see paper by R. Kelly, IEEE TAC, 1998; available as extra material on the course web)

Robotics 2 30
 Limits of robot regulation controllers
n response times needed for reaching the desired steady state are not
easily predictable in advance
n depend heavily on robot dynamics, on PD/PID gains, on the required

total displacement, and on the interested area of robot workspace

n integral term (when present) needs some time to “unload” itself from

the error history accumulated during transients

n large initial errors are stored in the integral term

n anti-windup schemes stop the integration when commands saturate

n … an intuitive explanation for the success of “saturated” PID law

n control efforts in the few first instants of motion typically exceed by far
those required at steady state
n especially for high positional gains

n may lead to saturation (hard nonlinearity) of robot actuators

Robotics 2 31
 Regulation in industrial robots
n in industrial robots, the planner generates a reference trajectory 𝑞¿ (𝑡)
even when the task requires only positioning/regulation of the robot
n “smooth” enough, with a user-defined transfer time 𝑇
n reference trajectory interpolates initial and final desired position
𝑞¿ (0) = 𝑞(0) 𝑞¿ (𝑡 ≥ 𝑇) = 𝑞|
n 𝑞¿ (𝑡) is used within a control law of the form
e.g., PD with
𝑢 = 𝐾ƒ 𝑞¿ (𝑡) − 𝑞 + 𝐾„ 𝑞̇ ¿ (𝑡) − 𝑞̇ + 𝑔(𝑞) gravity
cancellation
often neglected
n in this way, the position error is initially zero
n robot motion stays only “in the vicinity” of the reference trajectory
until 𝑡 = 𝑇, typically with small position errors (gains can be larger!)
n final regulation is only a “local” problem (𝑒 𝑇 = 𝑞| − 𝑞(𝑇) is small)
Robotics 2 32
 Qualitative comparison
n no saturation of commands: in principle, much larger gains can be used
n better prediction of settling times: local exponential convergence (designed
on the linear approximation of the robot dynamics around (𝑞| , 0))
n “fine tuning” of control gains is easier, but still a tedious and delicate task

𝑞𝑑 step 𝑢(0)
variation
of desired
position
𝑢| = 𝑔(𝑞| )
𝑞(0)
joint variables time- control commands
varying
𝑞𝑑 position
reference
𝑞¿ (𝑡) 𝑢(0)
𝑢| = 𝑔(𝑞| )
𝑞(0)
𝑇 = planned
final time
Robotics 2 33
 Quantitative comparison
planar 2R arm

robot data: links are uniform thin rods

no gravity (horizontal plane)

rest-to-rest motion task:

𝑞(0) = (0, 0) (straight) → 𝑞| = (𝜋/3, 𝜋/2)

interpolating trajectory: cubic polynomials

three case studies
a) low gains (overdamped) 𝐾ƒ = diag 80, 40 , 𝐾„ = diag 60, 30
vs interpolating trajectory in 𝑇 = 2 s
b) medium gains (very overdamped) 𝐾ƒ = diag 200, 100 , 𝐾„ = diag 200, 100
vs interpolating trajectory in 𝑇 = 2 s
c) high gains 𝐾ƒ = diag 1250, 180 , 𝐾„ = diag 200, 70
vs interpolating trajectory in 𝑇 = 1 s, with torque saturation 𝑢#,•–— = 30 Nm,
𝑢$,•–— = 10 Nm
Robotics 2 34
 Comparison on a planar 2R arm – case a

PD with low gains

𝐾ƒ = diag 80, 40
𝐾„ = diag 60, 30
(overdamped)
a reduction of the
peak control effort
by a factor of 100
on joint 1 &
by a factor of 30
on joint 2
max torques
PD with same gains of 7 and 2.3 Nm
on interpolating
trajectory of 𝑇 = 2 s

Robotics 2 35
 Comparison on a planar 2R arm – case b

PD with medium gains

𝐾ƒ = diag 200, 100
𝐾„ = diag 200, 100 even stronger
(very overdamped) peak reduction,
with similar total
control effort,
plus improved
tracking of
reference trajectory
on both joints

max torques
PD with same gains of 7.5 and 2.4 Nm
on interpolating
trajectory of 𝑇 = 2 s

Robotics 2 36
 Comparison on a planar 2R arm – case c

PD with high gains

𝐾ƒ = diag 1250, 180
𝐾„ = diag 200, 70
position overshoot
and long saturations
are avoided,
torque saturation with very good
𝑢#,•–— = 30 Nm tracking of the
𝑢$,•–— = 10 Nm faster reference
trajectory
max torques
PD with same gains of 30 and 9.5 Nm
on interpolating
trajectory of 𝑇 = 1 s

Robotics 2 37